TL;DR:
OpenAI's customizable GPTs have incredible artistic expression but suffer when returning structured, validated inputs.
Our Bots
Despite some limitations OpenAI's (#openai ) new create #gpts are incredible. Justin Lee and I have been playing around with creating GPTs this past week and the results have been remarkable.
We created two bots:
1. Logogen: Creating a good custom logos is costly and time consuming for new businesses and usually initial logos are modified anyway. Logogen fixes this by creating an agent that can create custom logos for you using rough sketches and text (see video). (Note: idea came from previous work with Adrian Ziegler!)
2. Unbiased: Given the uncertainty and speculation that typically surrounds elections most polls, news history, broadcasted results are not accurate; having all the facts, learning about candidates and propositions through questions, and knowing when certain facts are biased (left leaning, right leaning because of sampling bias for instance) could help people make better decisions. Unbiased fixes this by enforcing that all produced content is validated or sourced and explaining why a certain source may be biased.
Lessons
We learned two important properties about OpenAI's agents today (like literally the moment I'm writing this b/c OpenAI is moving so fast that these properties might change tomorrow)
1. They're really good at artistic interpolation meaning given multiple ideas ChatGPT can produce good generalizations of these ideas. But, by definition, art is really only related by semantics (two pictures of a t-rex can look significantly different but still be a t-rex). This property means that there's a large set of "good" outcomes that can be produced if I ask Logogen "give me a t-rex logo".
2. On the other hand factual look-ups are not yet rigorously structured in the GPT agents I played with. If I ask Unbiased who's leading the presidential election polls then it will search Bing for an answer and the first result that roughly matches my query will be returned. This makes sense since Unbiased doesn't have the bandwidth to return all polls on the internet (*cough yet *cough) which means that more human validated approaches for fetching factual data are needed. IMO there are three hypothetical options to solve this:
a) Websites (or OpenAI) start maintaining vector DB stores that allow for vector searches.
b) Agents like Unbiased are restricted to human validated APIs that can fetch facts and will have known biases.
c) The GPT creator specifies a set of endpoints (websites and APIs) and maintains them (forcing a human in the loop).
Right now we're at C (kinda like when GPT4 didn't upload images).
All in all I'm excited to see where #gptstore goes, as for logo and brand creators it looks like OpenAI is coming after your jobs and once it starts coding its own models it’ll start coming after mine ahaha.